Source code for deepmreye.util.model_opts

from . import util




def get_opts():
    opts = {}

    opts["kernel"] = 3  # Kernel size for all convolutional layers
    opts["lr"] = 0.00002  # Learning rate
    opts["filters"] = 32  # Number of filters in convolutional layers
    opts["multiplier"] = 2  # How much does the number of filters increase in each layer
    opts["depth"] = 4  # Maximum number of layers
    opts["dropout_rate"] = 0.1  # Dropout ratio for fully connected layers
    opts["num_dense"] = 2  # Number of fully connected layers
    opts["num_fc"] = 1024  # Number of units in fully connected layer
    opts["gaussian_noise"] = 0  # How much gaussian noise is added (unit = standard deviation)
    opts["activation"] = util.mish  # Activation function for all layers
    opts["groups"] = 8  # Number of groups to normalize across (see GroupNorm)
    opts["inner_timesteps"] = 10  # Default number of subTR samples which are being reconstructed

    # Loss weights
    opts["loss_euclidean"] = 1  # Loss weight for euclidean distance
    opts["loss_confidence"] = 0.1  # Loss weight for uncertainty measure

    # Training
    opts["epochs"] = 25  # Number of epochs
    opts["steps_per_epoch"] = 1500  # Number of steps per training epoch
    opts["validation_steps"] = 1500  # Number of steps per validation epoch
    opts["train_test_split"] = 0.6  # Default proportion of train (60%)-test(40%) split
    opts["batch_size"] = 8  # Batch size used for training the model
    opts["mixed_batches"] = True  # If true, each batch contains samples across participants
    opts["mc_dropout"] = False  # If true, monte carlo dropout is used
    opts["rotation_x"] = 5  # Augmentation parameter, rotation in x-axis
    opts["rotation_y"] = 5  # Augmentation parameter, rotation in y-axis
    opts["rotation_z"] = 5  # Augmentation parameter, rotation in z-axis
    opts["shift"] = 4  # Augmentation parameter, shift in all axes
    opts["zoom"] = 0.15  # Augmentation parameter, zoom in all axes

    return opts